A PPG Signal Denoising and Reconstruction Method Based on Causal Disentanglement Network

Abstract

The present invention proposes a method for denoising and reconstructing PPG signals based on a causal decoupling network. The method includes: inputting the original PPG signal into a signal preprocessing module for preprocessing; inputting the preprocessed feature vector into the causal decoupling network module to decouple the signal features using a Transformer encoder and a latent factor projector; based on the signal reconstruction module, decoding and reconstructing the decoupled latent factors using a multi-layer perceptron decoder to obtain the reconstructed PPG signal. By innovatively applying causal decoupling technology, the present invention decomposes the mixed latent representation of the PPG signal into multiple independent physiological factors, can effectively separate and remove the interference of complex noises such as motion artifacts and ambient light, and at the same time maximally retains key physiological information such as heart rate and respiration, significantly improving the signal-to-noise ratio, feature separability and overall quality of the signal.

Description

Technical Field

[0001] The present invention relates to the technical field of biomedical signal processing, and particularly to a method for denoising and reconstructing PPG signals based on a causal disentanglement network. Background Art

[0002] The PPG signal is a non-invasive physiological signal that measures the change in human blood volume through a photoelectric sensor, and is widely used in fields such as heart rate monitoring, blood pressure estimation, and blood oxygen saturation measurement. The accuracy and reliability of the PPG signal are crucial for medical diagnosis and health monitoring. However, the PPG signal is easily interfered by various noises during the acquisition process, such as ambient light noise, motion artifacts, device noise, etc. These noises will reduce the signal quality and affect subsequent physiological parameter estimation and disease diagnosis.

[0003] Traditional PPG signal processing methods mainly rely on techniques such as filter design and signal smoothing to remove noise. However, these methods can often only process specific types of noise, and may lose important physiological information while removing noise. In recent years, with the rapid development of deep learning technology, signal processing methods based on deep learning have gradually received attention. Deep learning models, such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), have been used for denoising and feature extraction of PPG signals. However, most of these methods focus on the processing of a single type of noise, and have limited effects on PPG signal processing in complex noise environments.

[0004] In addition, the PPG signal contains information on multiple physiological factors, such as heart rate, respiration, vascular status, etc. These factors are intertwined, making the feature extraction and analysis of the signal more complex. How to effectively separate these physiological factors and retain important physiological information while denoising is an important challenge faced by the current PPG signal processing field.

[0005] In recent years, significant progress has been made in the field of causal disentanglement technology in image processing and speech signal processing. The goal of causal disentanglement is to decompose the input signal into multiple independent latent factors, and each factor corresponds to a specific physical or physiological process. In this way, noise can be removed more effectively while retaining the key information in the signal. However, there has been no research applying causal disentanglement technology to PPG signal processing yet. Summary of the Invention

[0006] In view of the above situation, the main purpose of the present invention is to propose a method for denoising and reconstructing PPG signals based on a causal disentanglement network to solve the above technical problems.

[0007] The present invention proposes a method for denoising and reconstructing PPG signals based on a causal decoupling network. The method includes the following steps:

[0008] Step 1: Construct a causal decoupling network module based on the Transformer architecture, and construct a signal reconstruction module based on a signal reconstructor. The signal preprocessing module, the causal decoupling network module, the signal reconstruction module, and the model training module constitute a reconstruction model;

[0009] Among them, the causal decoupling network module includes a Transformer encoder and a latent factor projector. The Transformer encoder includes a multi-layer perceptron and a feed-forward network. The signal reconstruction module includes a multi-layer perceptron decoder;

[0010] Step 2: Input the original PPG signal into the signal preprocessing module for preprocessing to obtain a preprocessed feature vector;

[0011] Step 3: Input the preprocessed feature vector into the causal decoupling network module to perform signal feature decoupling using the Transformer encoder and the latent factor projector to obtain decoupled latent factors;

[0012] Step 4: Based on the signal reconstruction module, use the multi-layer perceptron decoder to decode and reconstruct the decoupled latent factors to obtain a reconstructed PPG signal;

[0013] Step 5: Based on the model training module, construct a decoupling loss according to the decoupled latent factors, and construct a reconstruction loss according to the original PPG signal and the reconstructed PPG signal;

[0014] Weight and combine the joint decoupling loss and the reconstruction loss, and input them into the Adam optimizer to update the parameters of the reconstruction model to obtain an updated reconstruction model;

[0015] Obtain the final reconstruction result based on the updated reconstruction model.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0017] 1. By innovatively applying the causal decoupling technology, the present invention decomposes the mixed latent representation of the PPG signal into multiple independent physiological factors, which can effectively separate and remove the interference of complex noises such as motion artifacts and ambient light, while maximizing the retention of key physiological information such as heart rate and respiration, significantly improving the signal-to-noise ratio, feature separability, and overall quality of the signal;

[0018] 2. By adopting a context-aware Transformer architecture and combining it with conditional layer normalization technology, the present invention can effectively capture the long-range temporal dependencies and dynamic change characteristics in PPG signals. Meanwhile, external context conditions such as demographics and treatment information are incorporated for adaptive processing, improving the accuracy of feature representation and the robustness to signal changes under different individuals and states.

[0019] 3. By jointly optimizing the reconstruction loss and the decoupling loss function, the present invention not only ensures a high degree of consistency (high fidelity) between the reconstructed signal and the original clean signal, but also forcibly promotes the mutual independence of the latent representations of various physiological factors, which is conducive to more in-depth, reliable, and interpretable analysis of specific physiological components (such as heart rate variability, respiratory modulation, etc.) subsequently, thereby improving the accuracy of downstream applications (such as disease diagnosis and health status monitoring).

[0020] Additional aspects and advantages of the present invention will be partly given in the following description, partly will become apparent from the following description, or can be understood through the embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a flowchart of the steps of the PPG signal denoising and reconstruction method based on a causal decoupling network proposed by the present invention.

[0022] Figure 2 It is an overall framework diagram of the PPG signal denoising and reconstruction method based on a causal decoupling network proposed by the present invention.

[0023] Figure 3 It is a framework diagram of the signal preprocessing module of the PPG signal denoising and reconstruction method based on a causal decoupling network proposed by the present invention.

[0024] Figure 4 It is an architecture diagram of the causal decoupling module of the PPG signal denoising and reconstruction method based on a causal decoupling network proposed by the present invention.

[0025] Figure 5 It is an architecture diagram of the context-aware encoder of the PPG signal denoising and reconstruction method based on a causal decoupling network proposed by the present invention.

[0026] Figure 6 It is an architecture diagram of the latent factor decomposition of the PPG signal denoising and reconstruction method based on a causal decoupling network proposed by the present invention.

[0027] Figure 7 It is an architecture diagram of the signal reconstruction module of the PPG signal denoising and reconstruction method based on a causal decoupling network proposed by the present invention.

[0028] Figure 8This is the architecture diagram of the model training module for the PPG signal denoising and reconstruction method based on the causal decoupling network proposed by the present invention. Detailed implementation manners

[0029] The embodiments of the present invention will be described in detail below. Examples of the embodiments are shown in the drawings, wherein the same or similar reference numerals are the same or similar elements or elements having the same or similar functions from beginning to end. The embodiments described below with reference to the drawings are exemplary and are only used to explain the present invention and should not be construed as a limitation of the present invention.

[0030] Referring to the following description and drawings, these and other aspects of the embodiments of the present invention will be clear. In these descriptions and drawings, some specific implementation manners in the embodiments of the present invention are specifically disclosed as some ways to implement the principles of the embodiments of the present invention, but it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0031] Please refer to Figure 1 , an embodiment of the present invention proposes a PPG signal denoising and reconstruction method based on a causal decoupling network. The method includes the following steps:

[0032] Step 1: Construct a causal decoupling network module based on the Transformer architecture, construct a signal reconstruction module based on a signal reconstructor, and a signal preprocessing module, a causal decoupling network module, a signal reconstruction module, and a model training module form a reconstruction model;

[0033] Among them, the causal decoupling network module includes a Transformer encoder and a latent factor projector. The Transformer encoder includes a multi-layer perceptron and a feed-forward network. The signal reconstruction module includes a multi-layer perceptron decoder.

[0034] Step 2: Input the original PPG signal into the signal preprocessing module for preprocessing to obtain a preprocessed feature vector.

[0035] Please refer to Figure 2 and Figure 3 , in step 2, input the original PPG signal into the signal preprocessing module for preprocessing to obtain a preprocessed feature vector, which specifically includes the following steps:

[0036] Input the original PPG signal into the signal preprocessing module, and use one-dimensional convolution to divide the original PPG signal to obtain divided local feature vectors;

[0037] Add position encoding to obtain the added position encoding;

[0038] Perform an addition calculation on the divided local feature vectors and the added position encoding to obtain local feature vectors after adding the position encoding.

[0039] The original PPG signal is input into the signal preprocessing module, and one-dimensional convolution is used to divide the original PPG signal to obtain the divided local feature vectors. The relational expressions in the corresponding process are as follows:

[0040] ;

[0041] Among them, represents the divided local feature vectors, represents the result after one-dimensional convolution operation, represents the original PPG signal, represents the convolution kernel size, represents the stride;

[0042] In the step of adding positional encoding to obtain the added positional encoding, the relational expressions in the corresponding process are as follows:

[0043] ;

[0044] Among them, represents the positional encoding value at position and the even feature dimension index ; represents the sine function, represents the exponential part, represents the positional encoding value at position and the odd feature dimension ; represents the cosine function, represents the feature dimension index, represents the positional encoding function, represents the position index, represents the model dimension;

[0045] It should be noted that is used to scale the frequency.

[0046] In the step of adding the divided local feature vectors and the added positional encoding to obtain the local feature vectors with added positional encoding, the relational expressions in the corresponding process are as follows:

[0047] ;

[0048] Among them, represents the local feature vectors with added positional encoding.

[0049] Step 3: Input the preprocessed feature vectors into the causal decoupling network module, and use the Transformer encoder and the latent factor projector to decouple the signal features, obtaining the decoupled latent factors.

[0050] Please refer to Figure 4 、 Figure 5 and Figure 6 , in Step 3, input the preprocessed feature vectors into the causal decoupling network module, and use the Transformer encoder and the latent factor projector to decouple the signal features, obtaining the decoupled latent factors. The specific steps are as follows:

[0051] S101: Input the external condition information and use a multi-layer perceptron to encode the context information to generate the final conditional vector;

[0052] S102: Input the local feature vectors added with positional encoding into the Transformer encoder for tensor representation to obtain the first represented tensor;

[0053] S103: Perform conditional layer normalization on the first represented tensor and the final conditional vector to obtain the output of the first conditional layer normalization;

[0054] S104: Perform a multi-head self-attention mechanism on the output of the first conditional layer normalization to obtain the output of the multi-head self-attention mechanism;

[0055] S105: Perform regularization on the output of the multi-head self-attention mechanism to obtain the regularization result of the output of the multi-head self-attention mechanism. Apply a residual connection to the regularization result of the output of the multi-head self-attention mechanism and the first represented tensor to obtain the output tensor after being processed by the self-attention sublayer and applying the residual connection;

[0056] S106: Perform tensor representation on the output tensor after being processed by the self-attention sublayer and applying the residual connection to obtain the second represented tensor;

[0057] S107: Perform conditional layer normalization on the second represented tensor and the final conditional vector to obtain the output of the second conditional layer normalization;

[0058] S108: Perform mechanism processing on the output of the second conditional layer normalization using a feed-forward network to obtain the feed-forward network output;

[0059] S109: Perform regularization on the feed-forward network output to obtain the regularization result of the feed-forward network output. Apply a residual connection to the regularization result of the feed-forward network output and the second represented tensor to obtain the output tensor after being processed by the feed-forward network sublayer and applying the residual connection;

[0060] S110. Use the output tensor after being processed by the feed-forward network sub-layer and applying the residual connection as the input, and repeat the steps of S102 to S109 in an iterative manner to obtain the final output tensor;

[0061] S111. Perform a final conditional normalization operation and an average pooling operation on the final output tensor in sequence to obtain a mixed latent representation;

[0062] S112. Perform a latent factor decomposition process on the mixed latent representation to obtain a decomposition result;

[0063] S113. Use a multi-layer perceptron to extract shared features from the decomposition result to obtain the extracted features;

[0064] S114. Project the extracted features using a latent factor projector to obtain a list of output factors.

[0065] Input external conditional information is encoded for context information using a multi-layer perceptron to generate a final conditional vector. The relational expressions existing in the corresponding process are as follows:

[0066] ;

[0067] Among them, represents the encoded demographic and treatment-related information, represents a small multi-layer perceptron for encoding context information, represents a concatenation operation, represents treatment-related features, represents demographic-related features, represents the embedding vector of the state index, represents an embedding layer, represents the index of the state, represents the final conditional vector;

[0068] In the step of inputting the local feature vector after adding the position encoding into the Transformer encoder for tensor representation to obtain the first represented tensor, the relational expressions existing in the corresponding process are as follows:

[0069] ;

[0070] Among them, represents the input tensor before entering the self-attention sub-layer, represents the local feature vector after adding the position encoding as the input ;

[0071] In the step of performing conditional layer normalization on the first represented tensor and the final conditional vector to obtain the output of the first conditional layer normalization, the relational expressions existing in the corresponding process are as follows:

[0072] ;

[0073] Among them, represents the output of the first conditional layer normalization, represents being processed by conditional layer normalization;

[0074] In the step of performing the multi-head self-attention mechanism processing on the output of the first conditional layer normalization to obtain the output of the multi-head self-attention mechanism, the relational expressions existing in the corresponding process are as follows:

[0075] ;

[0076] Among them, represents the output of the multi-head self-attention mechanism, represents the multi-head self-attention operation, represents the query vector, represents the key vector, represents the value vector;

[0077] In the step of performing regularization processing on the output of the multi-head self-attention mechanism to obtain the regularization result of the output of the multi-head self-attention mechanism, and applying a residual connection to the regularization result of the output of the multi-head self-attention mechanism and the first represented tensor to obtain the output tensor after being processed by the self-attention sub-layer and applying the residual connection, the relational expressions existing in the corresponding process are as follows:

[0078] ;

[0079] Among them, represents the output tensor after being processed by the self-attention sub-layer and applying the residual connection, represents the regularization of the output of the multi-head self-attention mechanism.

[0080] Performing tensor representation on the output tensor after being processed by the self-attention sub-layer and applying the residual connection to obtain the second represented tensor, the relational expressions existing in the corresponding process are as follows:

[0081] ;

[0082] Among them, represents the input tensor before entering the feed-forward sub-layer;

[0083] In the step of performing conditional layer normalization on the second represented tensor and the final conditional vector to obtain the output of the second conditional layer normalization, the relational expressions existing in the corresponding process are as follows:

[0084] ;

[0085] Among them, represents the output of the second conditional layer normalization;

[0086] In the step of processing the output of the second conditional layer normalization using a feed-forward network to obtain the output of the feed-forward network, the relational expressions existing in the corresponding process are as follows:

[0087] ;

[0088] Among them, represents the output of the feed-forward network, represents being processed by the feed-forward network;

[0089] In the step of performing regularization processing on the output of the feed-forward network to obtain the regularization result of the output of the feed-forward network, and applying a residual connection to the regularization result of the output of the feed-forward network and the tensor represented by the second, the relational expressions existing in the corresponding process are as follows:

[0090] ;

[0091] Among them, represents the output tensor after being processed by the feed-forward network sub-layer and applying a residual connection, represents regularizing the output of the feed-forward network;

[0092] In the step of sequentially performing the final conditional normalization operation and average pooling operation on the final output tensor to obtain the hybrid latent representation, the relational expressions existing in the corresponding process are as follows:

[0093] ;

[0094] Among them, represents the output after passing through the Transformer encoder layer, represents the final normalization layer, represents the output of the last Transformer encoder layer, represents the hybrid latent representation, represents calculating the mean, represents calculating the mean on the second dimension of the tensor;

[0095] In the step of performing latent factor decomposition processing on the hybrid latent representation to obtain the decomposition result, the relational expressions existing in the corresponding process are as follows:

[0096] ; ;

[0097] Among them, represents the dimension of the conditional vector, represents the input of the shared multi-layer perceptron for generating latent factors;

[0098] In the step of using a multi-layer perceptron to extract shared features from the decomposition result to obtain the extracted features, the relational expressions in the corresponding process are as follows:

[0099] ;

[0100] Among them, represents the features extracted by the shared perceptron ;

[0101] In the step of projecting the extracted features using a latent factor projector to obtain the output factor sequence, the relational expressions in the corresponding process are as follows:

[0102] ;

[0103] Among them, represents the th latent factor, represents the th independent linear projection head, represents the output factor sequence, represents the 1st latent factor, represents latent factors.

[0104] Step 4: Based on the signal reconstruction module, use a multi-layer perceptron decoder to decode and reconstruct the decoupled latent factors to obtain the reconstructed PPG signal.

[0105] Please refer to Figure 7 , in Step 4, based on the signal reconstruction module, use a multi-layer perceptron decoder to decode and reconstruct the decoupled latent factors to obtain the reconstructed PPG signal, which specifically includes the following steps:

[0106] Perform splicing processing on the latent factors of the output factor sequence to obtain a spliced vector. The relational expressions in the corresponding process are as follows:

[0107] ;

[0108] Among them, represents the vector spliced from all latent factors;

[0109] Use a multi-layer perceptron to reconstruct the spliced vector to obtain the reconstructed PPG signal. The relational expressions in the corresponding process are as follows:

[0110] ;

[0111] Among them, represents the reconstructed PPG signal, represents the multi-layer perceptron.

[0112] It should be noted that is used as a decoder for reconstruction in this step.

[0113] Step 5: Based on the model training module, construct a decoupling loss according to the decoupled latent factors, and construct a reconstruction loss according to the original PPG signal and the reconstructed PPG signal;

[0114] Weight and combine the joint decoupling loss and the reconstruction loss, and input them into the Adam optimizer to update the parameters of the reconstruction model, obtaining the updated reconstruction model;

[0115] Obtain the final reconstruction result based on the updated reconstruction model.

[0116] Please refer to Figure 8 , in Step 5, construct a decoupling loss according to the decoupled latent factors, which specifically includes the following steps:

[0117] Based on the output factor sequence, obtain the vector concatenated by all latent factors, perform central factorization on the vector concatenated by all latent factors, and obtain all the centralized latent factors. The corresponding relationship in the process is as follows:

[0118] ;

[0119] Among them, represents the total dimension of all latent factors, represents the dimension of each latent factor, represents all the centralized latent factors, represents calculating the mean of the latent factors in the batch dimension ;

[0120] Calculate the covariance matrix of the centralized latent factors to obtain the covariance matrix of the latent factors. The corresponding relationship in the process is as follows:

[0121] ;

[0122] Among them, represents the covariance matrix of the latent factors, represents all the centralized latent factors transpose , represents the batch size;

[0123] Square the off - diagonal elements of the covariance matrix of the latent factors to obtain the decoupling loss. The relationship in the corresponding process is as follows:

[0124] ;

[0125] where, represents the decoupling loss, represents the operation of squaring the off - diagonal elements of the covariance matrix .

[0126] Construct the reconstruction loss based on the original PPG signal and the reconstructed PPG signal. The relationship in the corresponding process is as follows:

[0127] ;

[0128] where, represents the reconstruction loss, represents the number of output channels, represents the signal length, represents the square of solving the Frobenius norm, represents the Frobenius norm;

[0129] In the step of weighted combination of the joint decoupling loss and the reconstruction loss, the relationship of the total loss is as follows:

[0130] ;

[0131] where, represents the total loss, represents the weight of the reconstruction loss, represents the weight of the decoupling loss.

[0132] Weight and combine the joint decoupling loss and the reconstruction loss, and input them into the Adam optimizer to update the parameters of the reconstruction model. The relationship in the corresponding process is as follows:

[0133] ;

[0134] where, represents the parameters of the reconstruction model at the -th step, represents the learning rate, represents the gradient calculation of the reconstruction model parameters at the -th time step.

[0135] Furthermore, the update rule of the Adam optimizer is as follows:

[0136] ;

[0137] Among them, represents the first-order moment estimation of the gradient at the th step, represents the second-order moment estimation of the gradient at the th step, represents the gradient of the total loss function with respect to the reconstruction model parameter , and respectively represent the two decay rate parameters of the Adam optimizer, represents the numerical stability term, represents the first-order moment estimation after bias correction, represents the second-order moment estimation after bias correction, represents the bias correction term for the first-order moment estimation of the gradient at the th step, represents the bias correction term for the second-order moment estimation of the gradient at the th step.

[0138] It should be understood that each part of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application-specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0139] In the description of this specification, the description referring to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0140] The above-described embodiments merely represent several implementation manners of the present invention, and their descriptions are relatively specific and detailed, but should not be construed as limiting the scope of the patent of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention should be subject to the appended claims.

Claims

1. A PPG signal denoising and reconstruction method based on a causal decoupling network, characterized in that: The method comprises the following steps: Step 1: construct a causal decoupling network module based on the Transformer architecture, and construct a signal reconstruction module based on the signal reconstructor. The signal preprocessing module, the causal decoupling network module, the signal reconstruction module and the model training module constitute the reconstruction model; The causal decoupling network module includes a Transformer encoder and a latent factor projector, the Transformer encoder includes a multi-layer perceptron and a feedforward network, and the signal reconstruction module includes a multi-layer perceptron decoder; Step 2: Input the original PPG signal into the signal preprocessing module for preprocessing to obtain the preprocessed feature vector; Step 3: Input the preprocessed feature vector into the causal decoupling network module, use the Transformer encoder and latent factor projector to decouple the signal features, and obtain the decoupled latent factors; Step 4: Based on the signal reconstruction module, the decoupled latent factors are decoded and reconstructed using a multi-layer perceptron decoder to obtain a reconstructed PPG signal; Step 5: Based on the model training module, a decoupling loss is constructed according to the decoupled latent factors, and a reconstruction loss is constructed according to the original PPG signal and the reconstructed PPG signal; The joint decoupling loss and the reconstruction loss are weighted and combined, and input into the Adam optimizer to update the parameters of the reconstruction model to obtain the updated reconstruction model; The final reconstruction result is obtained based on the updated reconstruction model.

2. The PPG signal denoising and reconstruction method based on a causal decoupling network according to claim 1, characterized in that: In step 2, the original PPG signal is input into a signal preprocessing module for preprocessing to obtain a preprocessed feature vector, which specifically includes the following steps: The original PPG signal is input into the signal preprocessing module, and the original PPG signal is divided and processed using one-dimensional convolution to obtain the divided local feature vector; Add the position code to obtain the added position code; The divided local feature vector and the added position code are added to obtain the local feature vector after adding the position code.

3. The PPG signal denoising and reconstruction method based on a causal decoupling network according to claim 2, characterized in that: The original PPG signal is input into the signal preprocessing module, and the original PPG signal is divided and processed using one-dimensional convolution to obtain the local feature vector after division. The relationship between the corresponding process is as follows: ； in, represents the local feature vector after partitioning, It means that after one-dimensional convolution operation, represents the original PPG signal, represents the convolution kernel size, represents the step length; In the step of adding the position code and obtaining the added position code, the relationship between the corresponding process is as follows: ； in, Indicates at location and even-numbered feature dimension index The positional encoding value at represents the sine function, represents the exponential part, Indicates at location and odd feature dimensions The positional encoding value at represents the cosine function, represents the feature dimension index, represents the position encoding function, Represents the position index, Represents the model dimension; In the step of adding the divided local feature vector and the added position code to obtain the local feature vector after adding the position code, the corresponding process has the following relationship: ； in, Represents the local feature vector after adding position encoding.

4. The PPG signal denoising and reconstruction method based on a causal decoupling network according to claim 3 is characterized in that: In step 3, the preprocessed feature vector is input into the causal decoupling network module to perform signal feature decoupling using the Transformer encoder and the latent factor projector to obtain the decoupled latent factor, which specifically includes the following steps: S101, input external condition information and use a multi-layer perceptron to encode context information to generate a final condition vector; S102, inputting the local feature vector after adding the position encoding into the Transformer encoder for tensor representation to obtain a first represented tensor; S103, performing a conditional layer normalization operation on the first represented tensor and the final conditional vector to obtain a first conditional layer normalized output; S104, performing multi-head self-attention mechanism processing on the normalized output of the first conditional layer to obtain the output of the multi-head self-attention mechanism; S105, performing regularization processing on the output of the multi-head self-attention mechanism to obtain a regularized result of the output of the multi-head self-attention mechanism, applying a residual connection to the regularized result of the output of the multi-head self-attention mechanism and the first represented tensor, and obtaining an output tensor processed by the self-attention sub-layer and applying the residual connection; S106, performing a tensor representation on the output tensor after being processed by the self-attention sub-layer and applying the residual connection to obtain a second represented tensor; S107, performing a conditional layer normalization operation on the second represented tensor and the final conditional vector to obtain a second conditional layer normalized output; S108, using a feedforward network to process the normalized output of the second conditional layer to obtain a feedforward network output; S109, performing regularization processing on the feedforward network output to obtain a regularized result of the feedforward network output, applying a residual connection to the regularized result of the feedforward network output and the second represented tensor, to obtain an output tensor processed by the feedforward network sublayer and applying the residual connection; S110, taking the output tensor processed by the feedforward network sublayer and applying the residual connection as input, and iteratively repeating steps S102 to S109 to obtain a final output tensor; S111, performing final conditional normalization operation and average pooling operation on the final output tensor in sequence to obtain a mixed potential representation; S112, performing latent factor decomposition processing on the mixed latent representation to obtain a decomposition result; S113, extracting shared features from the decomposition results using a multi-layer perceptron to obtain extracted features; S114. Project the extracted features using a latent factor projector to obtain an output factor list.

5. The PPG signal denoising and reconstruction method based on a causal decoupling network according to claim 4, characterized in that: The input external condition information uses a multi-layer perceptron to encode the context information and generate the final condition vector. The corresponding process has the following relationship: ； in, represents the coded demographic and treatment related information, represents a small multilayer perceptron used to encode contextual information, Represents a splicing operation, Indicates characteristics relevant to treatment, represents characteristics related to demographics, represents the embedding vector of the state index, represents an embedding layer, The index representing the state, represents the final conditional vector; In the step of inputting the local feature vector after adding the position encoding into the Transformer encoder for tensor representation and obtaining the first represented tensor, the corresponding process has the following relationship: ； in, represents the input tensor before entering the self-attention sublayer, Represents the local feature vector after adding position encoding as input ; In the step of performing conditional layer normalization operation on the first represented tensor and the final conditional vector to obtain the normalized output of the first conditional layer, the corresponding process has the following relationship: ； in, represents the normalized output of the first conditional layer, Indicates that the conditional layer has been normalized; In the step of processing the normalized output of the first conditional layer by the multi-head self-attention mechanism to obtain the output of the multi-head self-attention mechanism, the corresponding process has the following relationship: ； in, represents the output of the multi-head self-attention mechanism, represents a multi-head self-attention operation, represents the query vector, represents the key vector, represents a value vector; In the step of regularizing the output of the multi-head self-attention mechanism to obtain the regularized result of the output of the multi-head self-attention mechanism, applying residual connection to the regularized result of the output of the multi-head self-attention mechanism and the first represented tensor, and obtaining the output tensor after being processed by the self-attention sub-layer and applying the residual connection, the relationship between the corresponding processes is as follows: ； in, represents the output tensor after being processed by the self-attention sublayer and applying the residual connection, Represents the output regularization of the multi-head self-attention mechanism.

6. The PPG signal denoising and reconstruction method based on a causal decoupling network according to claim 5, characterized in that: The output tensor after the self-attention sub-layer processing and residual connection is represented as a tensor to obtain the second tensor. The corresponding process has the following relationship: ； in, Represents the input tensor before entering the feed-forward sublayer; In the step of performing conditional layer normalization operation on the second represented tensor and the final conditional vector to obtain the normalized output of the second conditional layer, the corresponding process has the following relationship: ； in, represents the normalized output of the second conditional layer; In the step of using the feedforward network to process the normalized output of the second conditional layer and obtain the output of the feedforward network, the corresponding process has the following relationship: ； in, represents the output of the feedforward network, Indicates that it has been processed by a feedforward network; In the steps of regularizing the feedforward network output to obtain the regularized result of the feedforward network output, applying residual connection to the regularized result of the feedforward network output and the second represented tensor, and obtaining the output tensor after being processed by the feedforward network sublayer and applying the residual connection, the corresponding process has the following relationship: ； in, represents the output tensor after being processed by the feedforward network sublayer and applying the residual connection, Represents regularization of the feedforward network output; In the step of performing the final conditional normalization operation and average pooling operation on the final output tensor in sequence to obtain the mixed potential representation, the corresponding process has the following relationship: ； in, Indicates passing The output of the layer Transformer encoder, represents the final normalization layer, represents the output of the last layer of Transformer encoder, represents the mixed latent representation, represents the calculation of the mean, Indicates that the mean is calculated on the second dimension of the tensor; In the step of performing latent factor decomposition on the mixed latent representation to obtain the decomposition result, the corresponding process has the following relationship: ； ； in, represents the dimension of the conditional vector, represents the input of the shared multilayer perceptron used to generate latent factors; In the step of extracting shared features from the decomposition results using a multi-layer perceptron to obtain the extracted features, the corresponding process has the following relationship: ； in, Represented by shared perceptron Extracted features; In the step of projecting the extracted features using the latent factor projector to obtain the output factor sequence, the corresponding process has the following relationship: ； in, Indicates potential factors, Indicates Independent linear projection heads, represents the output factor sequence, represents the first latent factor, express potential factor.

7. The PPG signal denoising and reconstruction method based on a causal decoupling network according to claim 6, characterized in that: In step 4, based on the signal reconstruction module, the decoupled latent factors are decoded and reconstructed using a multi-layer perceptron decoder to obtain a reconstructed PPG signal, which specifically includes the following steps: The potential factors of the output factor sequence are concatenated to obtain the concatenated vector. The corresponding relationship in the process is as follows: ； in, A vector representing the concatenation of all latent factors; The concatenated vectors are reconstructed using a multi-layer perceptron to obtain the reconstructed PPG signal. The corresponding relationship in the process is as follows: ； in, represents the reconstructed PPG signal, Represents a multi-layer perceptron.

8. The PPG signal denoising and reconstruction method based on a causal decoupling network according to claim 7, characterized in that: In step 5, the decoupling loss is constructed according to the decoupled potential factors, which specifically includes the following steps: Based on the output factor sequence, the concatenated vector of all potential factors is obtained, and the concatenated vector of all potential factors is centrally factored to obtain all the centralized potential factors. The corresponding relationship in the process is as follows: ； in, represents the total dimension of all latent factors, represents the dimension of each latent factor, represents all potential factors after centering, In the batch dimension Calculate the mean of the latent factors on ; The covariance matrix of all the latent factors after centralization is calculated to obtain the covariance matrix of the latent factors. The relationship between the corresponding process is as follows: ； in, represents the covariance matrix of the latent factors, Represents all potential factors after centering Transpose , Indicates the batch size; The covariance matrix of the latent factors is squared for off-diagonal elements to obtain the decoupling loss. The corresponding process has the following relationship: ； in, represents the decoupling loss, Represents the covariance matrix Square the off-diagonal elements of .

9. The PPG signal denoising and reconstruction method based on a causal decoupling network according to claim 8, characterized in that: The reconstruction loss is constructed based on the original PPG signal and the reconstructed PPG signal. The relationship between the corresponding process is as follows: ； in, represents the reconstruction loss, Indicates the number of output channels, Indicates the signal length, represents solving the square of the Frobenius norm, represents the Frobenius norm; In the step of weighted combination of joint decoupling loss and reconstruction loss, the total loss is related as follows: ； in, represents the total loss, represents the weight of the reconstruction loss, Represents the weight of the decoupling loss.

10. The PPG signal denoising and reconstruction method based on a causal decoupling network according to claim 9, characterized in that: The joint decoupling loss and reconstruction loss are weighted and combined, and then input into the Adam optimizer to update the parameters of the reconstruction model. The corresponding process has the following relationship: ； in, Indicates The reconstruction model parameters of the step, represents the learning rate, Represents the reconstruction model parameters exist Gradient computation for a time step.

Images